小行星引力场在线学习与智能着陆控制

Asteroid exploration is of scientific significance, and the landing control technology is crucially one of technical challenges. The autonomy and intelligence of the landing controller is helpful future asteroid exploration missions in China. This project develops an intelligent autonomous landing control approach for asteroids with complex dynamical environments through the combination of traditional dynamics, control technologies and the newly developed deep learning, reinforcement learning. Two scientific problems are specifically addressed in this project. These are the online learning of the gravitational field in the process of spacecraft flying around and the issue of optimal real-time decision-making in the landing process. The specific research contents include building a new approximate model of the gravitational field using deep learning technology, the on-line estimation method and the learning algorithm for the gravitational field based on observation data of flight around , flight around orbit optimization, a new training architecture for intelligent controller based on trajectory optimization techniques and deep learning technology, constraint management algorithms and improvement strategy of terminal landing accuracy. This project is aimed at providing a fast and accurate gravitational field approximate model and a new intelligence landing control solution with real-time optimal decision-making ability for asteroid landing missions, and to reserve theoretical and technological foundation for future asteroid exploration activities in China.

小行星探测具有重大的科学意义，着陆控制技术是其中一项关键技术难题，提高着陆控制的自主性和智能化水平有助于我国未来小行星探测任务的顺利实施。本项目通过传统动力学、控制技术与深度学习、强化学习的创新结合，构建复杂环境下小行星智能自主着陆控制方案，并重点解决两个科学问题：探测器绕飞过程中的引力场在线学习问题，以及探测器着陆过程中的实时最优智能决策难题。具体研究内容包括基于深度学习技术搭建具有学习能力的引力场近似模型、基于绕飞观测数据的引力场反演及模型在线学习算法设计、绕飞轨道优化、综合轨迹优化技术和深度学习技术的智能控制器训练架构、辅助智能控制器的约束管理模块和着陆精度提升策略设计。旨在实现又快又准的引力场新模型和具有实时最优决策能力的智能着陆控制新技术，为我国未来的小行星探测任务提供理论和技术储备。

飞行器动力学与控制进一步发展面临新的棘手难题，近年来，人工智能技术基于存储、记忆、预训练的新模式为动力学与控制难题的解决提供了新可能。在本基金的支持下，申请人聚焦于复杂环境下小行星探测着陆控制任务，通过创新性融合经典动力学控制与深度学习、强化学习等技术，围绕数据驱动下动力学在线学习和智能自主控制开展工作，具体创新点包括: (1) 提出基础模型+网络模型+自适应项相结合的复合动力学模型构建方法，赋能模型在实时飞行数据驱动下的精度持续提升能力，探索了动力学建模新方式。(2) 提出了模型迭代学习和控制器迭代学习结合的双网络迭代学习架构，提升了控制器对飞行环境的动态匹配特性，并理论证明了网络学习对状态观测精度和控制精度的提升作用。（3）提出了新型Identifier-Actor-Critic网络学习架构，该架构区别于现有model-free的强化学习架构，而是将力学中的模型学习、控制中的最优控制数值求解与强化学习的思想结合起来，丰富了强化学习应用于航天控制的手段。在本基金的支持下，申请人以第一作者或通讯作者发表SCI论文13篇，其中Q1文章11篇。